Talking to a robot is much easier than to a person: how AI helps to learn new languages.
speakatalka.com is a project for learning foreign languages using an artificial intelligence. AI will help you learn new words, raise grammar and use the knowledge in the dialogue. In a live dialogue with hundreds of options for phrases.
Valentin Zykov, CEO of SpeakaTalka, tells about main feature of the app and why it will soon conquer the market of learning foreign languages.
What SpeakaTalka is all about?
Talking simply, this is an app that teaches you how to speak a foreign language using using the dialogues with AI. But it’s not that simple as it seems: you can learn words, grammars, build sentences in your head and still being affraid of speaking out loud. Or, if you talk, then just with the teacher. Being quiet with foreigners and thinking about your language barriers.
People aren’t shy of the AI and the dialogue doesn’t differ from the “real” one: AI adjusts to any phrase. The developers call it “student’s personal language environment”. This is the main advantage of the app — the possibility to have a dialogue in a foreign language without being tied to a native speaker or a teacher. AI not just speaks to the user, but also recognizes and points out the mistakes in pronunciation and helps to correct them.
First of all, Speakatalka helps to reduce the fear of speaking. Even if you know the grammar, but affraid to speak, it’s gonna be hard to talk to foreigners. I had it on my own. When I was in Thailand I was affraid to talk to locals. I saw how their faces were changing when I was trying to build a sentence.
How does it work?
The main vector of learning here is a big stroy, similar to a graphic novel. Each step the user makes detrermines how the dialogue will go further. So far there is just 1 story — “I am a big boss”. In the story, the user gets a job in an IT company. He will have to go through a difficult interview and convince the HR-manager that he is the one who deserves this position. A familiar story, isn’t it? So, the user won’t be just learn the necessary phrases in foreign language, but also will be mentally prepared for the real situations.
This story contains 15 lessons, each of them has its own vocabulary, grammar and dialogue itself. The prologue is same for everyone — even for the beginners. They will go through the basic grammar topics, learn how to build sentences. The more lessons you take, the harder they become and in the end the user will reach the Intermediate level and beyond! You don’t have to go through the whole story: you can take separate lessons. For example, learn articles.
Grammar and vocabulary lessons will always be FREE for users. To try the dialogues with AI you need to buy a subscription.
How the dialogue goes?
Once the user has completed the first 2 parts of the lesson — he will learn a few words and the topic from the grammar — he moves on to the dialogue. AI will tell the user how to make a sentence: he suggests the key words.
For example, it will ask: “Where are you going?”, and give hints: “to the right”, “library”. User answers: “I will go to the right because I want to go to the library.” The AI evaluates the response and generates the next phrase. This is how each lesson goes.
We need around 5,000–10,000 hours of speech divided in 5–15 sec clips to teach the AI indicate the voice correctly.
How the idea has become?
“During the work in other projects we used to talk to the Americans. Our team had to learn English to be able to speak. In the process of learning we found out that we have a lack of the live dialogues. That is how I came up with the idea to put AI in it” — Valentin says.
Before Speakatalka Valentin Zykov has worked for the other Startup — the platform where you were able to create the chat-bots and use it, for example, as the phone support. Unfortunately, this project was closed.
We had a synthesis, speech recognition, dialogue control, text classificator. We decided — why don’t we make a platform to learn the foreign languages.
There are 5 members in Speakatalka team: Valentin Zykov, Ivan Zykov, Maksim Riavin, Andrey Sergeev and Snezhana Zykova. They all know each other well since they are either siblings or friends which makes the teamwork more comfortable.
What difficulties Speakatalka have faced?
“The most difficult thing was the part of learning words and checking grammar, because it was necessary to repurpose the text recognition into the determining of the sounds the person pronounced incorrectly”, Valentin sums up. “It was necessary to to give hints and point to the word that user mispronounced. And then help to learn this word and show it in the context.”
The team has collected a database of 170,000 words from the open datasets. For each word, we chose from 5 to 15 examples of using them in context.
What stage of development in Speakatalka now?
By now, the project at the Alpha-testing stage. Some features can be found in socials. The grammar lessons and test dialogues are coming later this month. The team plans to launch the platfrom at the end of December.
Right now you can try the technology by pronuncing the phrase and take a look how clear you do it.
At the moment, datasets were made only for Russian, but the team collects data for 11 more languages: French, Spanish, Chinese and more…
Big Data is available is Speakatalka. The app suggests the user to pronounce the word or phrase in the mother-tongue and then in learned language. Pronounced things help to generate new datasets.
We plan to expand the platform and give the opportunity to learn 11 languages with help of AI. In the future we plan to get the advantage at the smaller markets.
Is this unique? Valentin has tested around 250 different apps, imcluding other start-ups. He says that he found only 1 app with similar functionality where the user learns the words, listens the dialogue and tries to pronounce it. Disadvantage of this dialogue is when the speaker says at least one word from the sentnce then the app marks it as correct answer. There was one more app but it was just with the correction of pronunciation.
Original publication is here
